Technology: Trying To Decipher Babel

The machine clearly does not like poetry. It won’t touch the stuff. Nor is it very fond of novels. Theoretically, it could cope with some of Hemingway’s short, simple sentences, though it could never make anything of long, convoluted passages from Faulkner. But give the Toshiba AS-TRANSAC computer a thoroughly dull, straightforward instruction manual, and it will earnestly chomp its way through page after page. What it does with those pages is the amazing part. The Toshiba machine has linguistic ability far beyond the powers of past generations of computers: it can translate, at least crudely, one language into another. In this case, the computer converts simple English into serviceable, if stilted Japanese.

The AS-TRANSAC is one of more than half a dozen machine-translation systems being energetically developed in Japan. With their strong thirst for information from other nations and a growing need to disseminate their documents around the world, the Japanese urgently require computers that can translate. A few machines, such as the Toshiba model and Fujitsu’s Atlas system, are already in operation, helping Japanese companies like Mazda translate technical material. A powerful computer called SHALT, designed by IBM Japan, is being used extensively for in-house translations. In 1988 SHALT converted four IBM manuals from English into Japanese. This year the target is 20 to 30. Predicts Kiyotaka Yasui, manager of the language and image- technology section at IBM Japan’s Tokyo research laboratory: “In five years the internal-publication department of IBM Japan will be fulfilling 100% of its translation requirements via machines.”

But human translators should have no fear that their jobs are imperiled — at least for now. None of the new systems are yet able to take a page of text and render it unerringly into a different language without the aid of a bilingual editor who can fine-tune the output for ambiguities in the ) vocabulary, to say nothing of shades of meaning. “A truly automatic system is a dream at the moment,” admits Makoto Ihara, manager of Toshiba’s computer product-planning department. Says Kazunori Muraki, a leading researcher at NEC: “Machine translation is only to reduce the work involved in human translation.”

And that it does. The present generation of machine-translation systems, which are priced between $30,000 and $70,000, can nearly double the output of translators of technical documents. The savings, especially for small firms unable to maintain a large staff of skilled translators, can be considerable.

“Seven or eight years ago,” says Koichi Takeda, a researcher at IBM Japan, “everyone was saying machine translation was a technology of the future. But now we have it.”

Considering the complexity of the task, the progress in machine translation has been startling. Essentially, the translating machine analyzes the syntax of an English sentence, determining its grammatical structure and identifying, for example, the subject, verb, objects and modifiers. Then the words are translated by an English-Japanese dictionary. Next, another part of the computer program analyzes the resulting awkward jumble of words and meanings, and generates an intelligible sentence based on the rules of Japanese syntax and the machine’s understanding of what the original English sentence meant.

That is not as simple as it sounds (assuming it sounds simple at all). Each computer company has devised strikingly different sets of programs to deal with the fiendish complexities of the two languages. One step in the IBM system, for example, refashions the English sentence structure and word order to resemble Japanese syntax. The result occasionally reads like the faulty work of a ninth-grade Japanese student of English. The articles and subjects are gone, and the verb dangles clumsily at the end. Only after the English sentence has been transformed into Japanese syntax are the words translated.

Japan’s computer makers are developing machines that can translate freely among several different languages. Fujitsu, for example, has a prototype called Atlas 2 that can deal with Japanese, French, German and English. In the near future, Spanish, Chinese and Korean will be added. To make such systems as simple as possible, programmers have invented a coded, largely numerical language called “interlingual.”

Instead of translating directly from Japanese into German, the computer ! would translate from Japanese into interlingual and then into German. This process cuts down on the number of dictionaries that programmers have to construct. A Japanese-interlingual dictionary would be needed, but not a Japanese-German, Japanese-French or Japanese-Spanish. Explains Hiroshi Uchida, a researcher at Fujitsu: “If we did not use interlingual, then each pair of languages would require the development of a specific set of grammatical rules and a bilingual dictionary. Interlingual acts as the hub of a wheel.”

The market for such machines will be vast. Says Yasuyo Kikuta, a researcher in artificial intelligence at Fujitsu: “Since we Japanese have so much trouble in the area of foreign languages, machine translation is the kind of tool all Japanese desire.” And since many people in other nations are not linguistic whizzes either, sales of the electronic translators should be brisk around the world.